White House Shut Down Anthropic's Models Over Suspected Chinese Access to Mythos, Anthropic Says It Was Never Told Why

Since the White House issued export controls on Anthropic's Fable 5 and Mythos last Friday and Anthropic took both models offline globally, the shutdown has been publicly framed around a jailbreak vulnerability. A new layer surfaced Saturday and Sunday: the administration may have had a second, more serious motive it never shared with Anthropic.

The China Suspicion

According to Semafor, which broke the story on June 13, the White House imposed export controls on Mythos partly because of suspicions that a China-linked group had gained access to the model. Semafor's Reed Albergotti cited a person familiar with the matter. The specific organization involved, how it got access, and how the administration learned about it are all unknown, Semafor reported.

The national security concern is straightforward. Anthropic itself has described Mythos as capable of autonomously identifying and exploiting software vulnerabilities. A capability dangerous enough that the company restricted access at launch to a small group of vetted partners, including the U.S. and U.K. governments, Amazon, Apple, and Google. If a China-linked actor got in, that capability could be studied and replicated through a process called model distillation, where a less capable AI is trained on a more powerful one's outputs to approximate its behavior.

The Verge's Terrence O'Brien noted on June 14 that this would not be Mythos's first unauthorized access problem. A Discord group reportedly had access to the model for two weeks before Anthropic discovered the breach and cut them off. Anthropic has previously acknowledged this fact.

China-US Focus, a publication affiliated with the China Institutes of Contemporary International Relations, published a June 4 piece by Li Yan framing Mythos as a geopolitical inflection point. Yan argued the model's capabilities mark a moment where AI power has concentrated into a few American firms, and called on China to accelerate its own development. That piece predates the shutdown and does not allege Chinese access. It makes clear, though, that Chinese government-linked analysts were closely watching Mythos long before the White House acted.

What Anthropic Was Actually Told

An Anthropic spokesperson told Semafor that the White House did NOT raise Chinese access to Mythos in its conversations with the company before imposing restrictions. The spokesperson said Anthropic prohibits use of its products from China.

If the administration's primary concern was a potential Chinese intelligence breach, and it chose not to tell the company being regulated about that concern, Anthropic's public pushback takes on a different character. The company called the shutdown order a misunderstanding and described the jailbreak evidence as verbal and narrow. The company may have been defending against the stated rationale while the actual rationale went unspoken.

David Sacks, Trump's AI and crypto adviser, posted on X on Saturday focusing entirely on the jailbreak angle. He alleged that when the administration notified Anthropic CEO Dario Amodei about the jailbreak vulnerability, Amodei dismissed it as not a serious risk and refused to fix it. Sacks said the export controls were issued reluctantly as a result. His post, per Semafor, made no mention of China.

The Strongest Counter-Argument

Anthropic's position deserves a fair hearing. The company's spokesperson says China was never brought up. Anthropic already bars Chinese access by policy. The jailbreak evidence, as Anthropic has characterized it across multiple statements this week, was described to them verbally and applied to a narrow use case, not the sweeping vulnerability the administration's public framing implied. If the White House had hard evidence of a Chinese breach, presenting it to Anthropic would have been the straightforward move. The fact that it apparently wasn't presented raises a legitimate question about whether the China concern was a confirmed finding or an unverified suspicion used to justify a broader action.

The White House has not confirmed the Semafor report. No intelligence assessment has been made public. No charges or regulatory findings have been announced.

What Remains Unknown

Four things are unresolved as of June 14. First, whether the suspected Chinese access was a confirmed breach or an unverified tip. Second, which organization is alleged to have accessed Mythos and how. Third, whether the jailbreak concern and the China concern are related or two separate reasons stacked on top of each other. Fourth, when, if ever, Fable 5 and Mythos come back online, and under what conditions.

Semafor reported that even Anthropic employees who are foreign nationals would have been barred from using the models under the original export control terms. A restriction so broad that Anthropic concluded pulling the models entirely was the only practical response.

Anthropic has not announced a timeline for restoring access, and the White House has not said what evidence would satisfy its concerns. Until one of those two things changes, both models stay dark.